def time_ccdcesque1(self, setup):
""" Bench with 'defaults' defined in setup with most tests turned off
"""
kwargs = version_kwargs(setup['kwargs'])
for i in range(n):
model = CCDCesque(**kwargs)
model.fit(setup['X'], setup['Y'], setup['dates'])
def pixel_CCDCesque(pipe, require, output, config=None):
""" Run :class:`yatsm.algorithms.CCDCesque` on a pixel
Users should pass to ``require`` both ``X`` and ``Y`` arguments, which
are interpreted as:
.. code-block:: python
X, Y = require[0], require[1:]
Args:
pipe (yatsm.pipeline.Pipe): Piped data to operate on
require (dict[str, list[str]]): Labels for the requirements of this
calculation
output (dict[str, list[str]]): Label for the result of this
calculation
config (dict): Configuration to pass to :class:`CCDCesque`. Should
contain `init` section
Returns:
yatsm.pipeline.Pipe: Piped output
"""
XY = pipe.data[require['data']].dropna('time', how='any')
X = XY[require['data'][0]]
Y = XY[require['data'][1:]].to_array()
model = CCDCesque(**config.get('init', {}))
model.py, model.px = Y.y, Y.x
model = model.fit(X, Y.values, XY['ordinal'])
pipe.record[output[RECORD][0]] = model.record
return pipe
def time_ccdcesque1(self, setup):
""" Bench with 'defaults' defined in setup with most tests turned off
"""
kwargs = version_kwargs(setup['kwargs'])
for i in range(n):
model = CCDCesque(**kwargs)
model.fit(setup['X'], setup['Y'], setup['dates'])
def time_ccdcesque3(self, setup):
""" Bench with remove_noise, dynamic_rmse turned on
"""
kwargs = version_kwargs(setup['kwargs'])
kwargs.update({'remove_noise': True, 'dynamic_rmse': True})
for i in range(n):
model = CCDCesque(**kwargs)
model.fit(setup['X'], setup['Y'], setup['dates'])
def time_ccdcesque2(self, setup):
""" Bench with remove_noise turned on
"""
kwargs = version_kwargs(setup['kwargs'])
kwargs.update({'remove_noise': True})
for i in range(n):
model = CCDCesque(**kwargs)
model.fit(setup['X'], setup['Y'], setup['dates'])
def time_ccdcesque1(self, setup):
""" Bench with 'defaults' defined in setup with most tests turned off
"""
kwargs = version_kwargs(setup['kwargs'])
model = CCDCesque(**kwargs)
for col in range(setup['Y'].shape[-1]):
_Y, _X, _dates = setup['Y'][..., col], setup['X'], setup['dates']
mask = np.in1d(_Y[-1, :], [0, 1])
model.fit(_X[mask, :], _Y[:, mask], _dates[mask])
def time_ccdcesque1(self, setup):
""" Bench with 'defaults' defined in setup with most tests turned off
"""
kwargs = version_kwargs(setup['kwargs'])
model = CCDCesque(**kwargs)
for col in range(setup['Y'].shape[-1]):
_Y, _X, _dates = setup['Y'][..., col], setup['X'], setup['dates']
mask = np.in1d(_Y[-1, :], [0, 1])
model.fit(_X[mask, :], _Y[:, mask], _dates[mask])
def time_ccdcesque3(self, setup):
""" Bench with remove_noise, dynamic_rmse turned on
"""
kwargs = version_kwargs(setup['kwargs'])
kwargs.update({'remove_noise': True, 'dynamic_rmse': True})
model = CCDCesque(**kwargs)
for col in range(setup['Y'].shape[-1]):
_Y, _X, _dates = setup['Y'][..., col], setup['X'], setup['dates']
mask = np.in1d(_Y[-1, :], [0, 1])
model.fit(_X[mask, :], _Y[:, mask], _dates[mask])
def time_ccdcesque2(self, setup):
""" Bench with remove_noise turned on
"""
kwargs = version_kwargs(setup['kwargs'])
kwargs.update({'remove_noise': True})
model = CCDCesque(**kwargs)
for col in range(setup['Y'].shape[-1]):
_Y, _X, _dates = setup['Y'][..., col], setup['X'], setup['dates']
mask = np.in1d(_Y[-1, :], [0, 1])
model.fit(_X[mask, :], _Y[:, mask], _dates[mask])
def _fetch_results_saved(self):
""" Read YATSM results and return """
self.yatsm_model = MockResult()
row, col = self.series[0].py, self.series[0].px
data_cfg = {
'output': os.path.join(self.location,
self.config['results_folder'].value),
'output_prefix': (self.config['results_pattern'].value
.replace('*', ''))
}
result_filename = get_output_name(data_cfg, row)
logger.info('Attempting to open: {f}'.format(f=result_filename))
if not os.path.isfile(result_filename):
qgis_log('Could not find result for row {r} ({fn})'.format(
r=row, fn=result_filename))
return
z = np.load(result_filename)
if 'record' not in z.files:
raise KeyError('Cannot find "record" within saved result ({})'
.format(result_filename))
if 'metadata' not in z.files:
raise KeyError('Cannot find "metadata" within saved result ({})'
.format(result_filename))
metadata = z['metadata'].item()
if 'design' not in metadata['YATSM']:
raise KeyError('Cannot find "design" within saved result metadata '
'({})'.format(result_filename))
self._design_info = metadata['YATSM']['design']
rec = z['record']
idx = np.where((rec['px'] == col) & (rec['py'] == row))[0]
self.yatsm_model.record = rec[idx]
def pixel_CCDCesque(work, require, output, **config):
""" Run CCDCesque on a pixel
"""
arr = work['data'][require['data']].dropna('time', how='any').to_array()
model = CCDCesque(**config.get('init', {}))
model.py, model.px = arr.y, arr.x
ordinal = arr.indexes['time'].map(lambda x: x.toordinal())
design = config.get('fit', {}).get('design', '1 + ordinal')
X = patsy.dmatrix(design,
data=arr,
eval_env=patsy.EvalEnvironment.capture())
work['record'][output['record'][0]] = model.fit(X, arr.values, ordinal)
return work
def _fetch_results_saved(self):
""" Read YATSM results and return """
self.yatsm_model = MockResult()
row, col = self.series[0].py, self.series[0].px
data_cfg = {
'output':
os.path.join(self.location, self.config['results_folder'].value),
'output_prefix':
(self.config['results_pattern'].value.replace('*', ''))
}
result_filename = get_output_name(data_cfg, row)
logger.info('Attempting to open: {f}'.format(f=result_filename))
if not os.path.isfile(result_filename):
qgis_log('Could not find result for row {r} ({fn})'.format(
r=row, fn=result_filename))
return
z = np.load(result_filename)
if 'record' not in z.files:
raise KeyError(
'Cannot find "record" within saved result ({})'.format(
result_filename))
if 'metadata' not in z.files:
raise KeyError(
'Cannot find "metadata" within saved result ({})'.format(
result_filename))
metadata = z['metadata'].item()
if 'design' not in metadata['YATSM']:
raise KeyError('Cannot find "design" within saved result metadata '
'({})'.format(result_filename))
self._design = metadata['YATSM']['design_matrix']
self._design_info = metadata['YATSM']['design']
rec = z['record']
idx = np.where((rec['px'] == col) & (rec['py'] == row))[0]
self.yatsm_model.record = rec[idx]
def _fetch_results_live(self):
""" Run YATSM and get results """
logger.debug('Calculating YATSM results on the fly')
# Setup design matrix, Y, and dates
self.X = patsy.dmatrix(
self.controls['design'].value, {
'x': self.series[0].images['ordinal'],
'sensor': self.series[0].sensor,
'pr': self.series[0].pathrow
})
self._design_info = self.X.design_info.column_name_indexes
self.Y = self.series[0].data.astype(np.int16)
self.dates = np.asarray(self.series[0].images['ordinal'])
mask = self.Y[self.config['mask_band'].value[0] - 1, :]
Y_data = np.delete(self.Y,
self.config['mask_band'].value[0] - 1,
axis=0)
# Mask out masked values
clear = np.in1d(mask, self.mask_values, invert=True)
valid = get_valid_mask(Y_data, self.config['min_values'].value,
self.config['max_values'].value).astype(np.bool)
clear *= valid
# Setup parameters
estimator = sklearn.linear_model.Lasso(alpha=20)
reg = self.controls['regression_type'].value
if hasattr(yatsm.regression, 'packaged'):
if reg in yatsm.regression.packaged.packaged_regressions:
reg_fn = yatsm.regression.packaged.find_packaged_regressor(reg)
try:
estimator = jl.load(reg_fn)
except:
logger.error('Cannot load regressor: %s' % reg)
else:
logger.debug('Loaded regressor %s from %s' % (reg, reg_fn))
else:
logger.error('Cannot use unknown regression %s' % reg)
else:
logger.warning(
'Using failsafe Lasso(lambda=20) from scikit-learn. '
'Upgrade to yatsm>=0.5.1 to access more regressors.')
kwargs = dict(
estimator=estimator,
test_indices=self.controls['test_indices'].value,
consecutive=self.controls['consecutive'].value,
threshold=self.controls['threshold'].value,
min_obs=self.controls['min_obs'].value,
min_rmse=(None if self.controls['enable_min_rmse'].value else
self.controls['min_rmse'].value),
screening_crit=self.controls['screen_crit'].value,
remove_noise=self.controls['remove_noise'].value,
dynamic_rmse=self.controls['dynamic_rmse'].value,
)
self.yatsm_model = CCDCesque(**version_kwargs(kwargs))
# Don't want to have DEBUG logging when we run YATSM
log_level = logger.level
logger.setLevel(logging.INFO)
if self.controls['reverse'].value:
self.yatsm_model.fit(np.flipud(self.X[clear, :]),
np.fliplr(Y_data[:, clear]),
self.dates[clear][::-1])
else:
self.yatsm_model.fit(self.X[clear, :], Y_data[:, clear],
self.dates[clear])
if self.controls['commit_test'].value:
self.yatsm_model.record = postprocess.commission_test(
self.yatsm_model, self.controls['commit_alpha'].value)
# if self.controls['robust_results'].value:
# self.coef_name = 'robust_coef'
# self.yatsm_model.record = postprocess.refit_record(
# self.yatsm_model, 'robust'
# else:
# self.coef_name = 'coef'
if self.config['calc_pheno'].value:
# TODO: parameterize band indices & scale factor
ltm = pheno.LongTermMeanPhenology()
self.yatsm_model.record = ltm.fit(self.yatsm_model)
# Restore log level
logger.setLevel(log_level)
class YATSMTimeSeries(timeseries_stacked.StackedTimeSeries):
""" Timeseries driver for CCDCesque algorithm implemented in YATSM
Requires a working installation of YATSM. For more information, visit
the [YATSM Github website](https://github.com/ceholden/yatsm).
This driver requires the following Python packages in addition to basic
TSTools package dependencies:
* [`scikit-learn`](http://scikit-learn.org/stable/)
* [`patsy`](https://patsy.readthedocs.org/en/latest/)
* [`yatsm`](https://github.com/ceholden/yatsm)
"""
description = 'YATSM CCDCesque Timeseries'
location = None
mask_values = np.array([2, 3, 4, 255])
has_results = True
# Driver configuration
config = OrderedDict((
('stack_pattern', ConfigItem('Stack pattern', 'L*stack')),
('date_index', ConfigItem('Date index', [9, 16])),
('date_format', ConfigItem('Date format', '%Y%j')),
('cache_folder', ConfigItem('Cache folder', 'cache')),
('results_folder', ConfigItem('Results folder', 'YATSM')),
('results_pattern', ConfigItem('Results pattern', 'yatsm_r*')),
('mask_band', ConfigItem('Mask band', [8])),
('min_values', ConfigItem('Min data values', [0])),
('max_values', ConfigItem('Max data values', [10000])),
('metadata_file_pattern', ConfigItem('Metadata file pattern',
'L*MTL.txt')),
('calc_pheno', ConfigItem('LTM phenology', False)),
))
# Driver controls
controls_title = 'YATSM Algorithm Options'
controls = OrderedDict((
('calculate_live', ConfigItem('Calculate live', True)),
('consecutive', ConfigItem('Consecutive', 5)),
('min_obs', ConfigItem('Min obs.', 16)),
('threshold', ConfigItem('Threshold', 4.0)),
('enable_min_rmse', ConfigItem('Use min RMSE?', True)),
('min_rmse', ConfigItem('Min RMSE', 100.0)),
('design', ConfigItem('Design', '1 + x + harm(x, 1)')),
('test_indices', ConfigItem('Test indices', np.array([2, 3, 4, 5]))),
('dynamic_rmse', ConfigItem('Dynamic RMSE', True)),
('screen_crit', ConfigItem('Screening crit value', 400.0)),
('remove_noise', ConfigItem('Remove noise', True)),
('reverse', ConfigItem('Reverse', False)),
('regression_type', ConfigItem('Regression type', 'sklearn_Lasso20')),
('robust_results', ConfigItem('Robust results', False)),
('commit_test', ConfigItem('Commission test', False)),
('commit_alpha', ConfigItem('Commission test alpha', 0.10)),
))
def __init__(self, location, config=None):
super(YATSMTimeSeries, self).__init__(location, config=config)
# Check for YATSM imports
if not has_yatsm:
raise ImportError(has_yatsm_msg)
if self.config['calc_pheno'].value and not has_yatsm_pheno:
raise ImportError(has_yatsm_pheno_msg)
# Find extra metadata
self._init_metadata()
# Setup YATSM
self.yatsm_model = None
self.X = None
self.Y = None
self.coef_name = 'coef'
# Setup min/max values
desc, _min_values = self.config['min_values']
if len(_min_values) == 1:
_min_values = np.repeat(_min_values, self.series[0].count - 1)
self.config['min_values'] = ConfigItem(desc, _min_values)
desc, _max_values = self.config['max_values']
if len(_max_values) == 1:
_max_values = np.repeat(_max_values, self.series[0].count - 1)
self.config['max_values'] = ConfigItem(desc, _max_values)
def set_custom_controls(self, values):
logger.debug('Setting custom values')
for val, attr in zip(values, self.controls):
desc, current_val = self.controls[attr]
if isinstance(val, type(current_val)):
self.controls[attr] = ConfigItem(desc, val)
else:
# Make an exception for minimum RMSE since we can pass None
if attr == 'min_rmse' and isinstance(val, float):
self.controls[attr] = ConfigItem(desc, val)
else:
msg = 'Could not set {k} to {v} (current: {c})'.format(
k=attr, v=val, c=current_val)
raise ValueError(msg)
def fetch_results(self):
""" Read or calculate results for current pixel """
if self.controls['calculate_live'].value:
self._fetch_results_live()
else:
self._fetch_results_saved()
# Update multitemporal screening metadata
if self.yatsm_model:
if (self.controls['calculate_live'] and
hasattr(self.yatsm_model, 'X')):
self.series[0].multitemp_screened = \
np.in1d(self.X[:, 1], self.yatsm_model.X[:, 1],
invert=True).astype(np.uint8)
if self.config['calc_pheno'].value:
for rec in self.yatsm_model.record:
# Find dates in record
idx = np.where(
(self.series[0].images['ordinal'] >= rec['start']) &
(self.series[0].images['ordinal'] <= rec['end']))[0]
# Put observations into SPR/SUM/AUT
_spr = np.where(self.series[0].images['doy'][idx] <=
rec['spring_doy'])[0]
_sum = np.where((self.series[0].images['doy'][idx] >
rec['spring_doy']) &
(self.series[0].images['doy'][idx] <
rec['autumn_doy']))[0]
_aut = np.where(self.series[0].images['doy'][idx] >=
rec['autumn_doy'])[0]
self.series[0].pheno[idx[_spr]] = 'SPR'
self.series[0].pheno[idx[_sum]] = 'SUM'
self.series[0].pheno[idx[_aut]] = 'AUT'
def get_prediction(self, series, band, dates=None):
""" Return prediction for a given band
Args:
series (int): index of Series used for prediction
band (int): index of band to return
dates (iterable): list or np.ndarray of ordinal dates to predict; if
None, predicts for every date within timeseries (default: None)
Returns:
iterable: sequence of tuples (1D NumPy arrays, x and y) containing
predictions
"""
if series > 0:
return
if self.yatsm_model is None or len(self.yatsm_model.record) == 0:
return
if band >= self.yatsm_model.record[self.coef_name].shape[2]:
logger.debug('Not results for band %i' % band)
return
# Setup output
mx = []
my = []
# Don't predict with any categorical information
eqn = (self.controls['design'].value
if self.controls['calculate_live'].value
else self._design)
design = re.sub(r'[\+\-][\ ]+C\(.*\)', '', eqn)
coef_columns = []
for k, v in self._design_info.iteritems():
if not re.match('C\(.*\)', k):
coef_columns.append(v)
coef_columns = np.sort(np.asarray(coef_columns))
for rec in self.yatsm_model.record:
# Check for reverse
if rec['end'] < rec['start']:
i_step = -1
else:
i_step = 1
# Date range to predict
if dates is not None:
end = max(rec['break'], rec['end'])
_mx = dates[np.where((dates >= rec['start']) &
(dates <= end))[0]]
else:
_mx = np.arange(rec['start'], rec['end'], i_step)
if _mx.size == 0:
continue
# Coefficients to use for prediction
_coef = rec[self.coef_name][coef_columns, band]
# Setup design matrix
_mX = patsy.dmatrix(design, {'x': _mx}).T
# Predict
_my = np.dot(_coef, _mX)
# Transform ordinal back to datetime for plotting
_mx = np.array([dt.fromordinal(int(_x)) for _x in _mx])
mx.append(_mx)
my.append(_my)
return mx, my
def get_breaks(self, series, band):
""" Return break points for a given band
Args:
series (int): index of Series for prediction
band (int): index of band to return
Returns:
iterable: sequence of tuples (1D NumPy arrays, x and y) containing
break points
"""
if self.yatsm_model is None:
return
# Setup output
bx = []
by = []
if len(self.yatsm_model.record) > 0:
for rec in self.yatsm_model.record:
if rec['break'] != 0:
_bx = dt.fromordinal(int(rec['break']))
index = np.where(self.series[series].images['date'] ==
_bx)[0]
if (index.size > 0 and
index[0] < self.series[series].data.shape[1]):
bx.append(_bx)
by.append(self.series[series].data[band, index[0]])
else:
logger.warning('Could not determine breakpoint')
return bx, by
def get_residuals(self, series, band):
""" Return model residuals (y - predicted yhat) for a given band
Args:
series (int): index of Series for residuals
band (int): index of band to return
Returns:
iterable: sequence of tuples (1D NumPy arrays, x and y) containing
residual dates and values
"""
if self.yatsm_model is None:
return
rx, ry = [], []
X, y = self.get_data(series, band, mask=settings.plot['mask'])
predict = self.get_prediction(series, band, dates=X['ordinal'])
if predict is None:
return
date, yhat = predict
for _date, _yhat in zip(date, yhat):
idx = np.in1d(X['date'], _date)
resid = y[idx] - _yhat
rx.append(_date)
ry.append(resid)
return rx, ry
def get_plot(self, series, band, axis, desc):
""" Plot some information on an axis for a plot of some description
Args:
series (int): index of Series for residuals
band (int): index of band to return
axis (matplotlib.axes._subplots.Axes): a matplotlib axis to plot on
desc (str): description of plot, usually a plot class from
`tstools.plots`
Returns:
iterable: list of artists to include in legend
"""
artists = []
if desc == 'TSPlot':
for rec in self.yatsm_model.record:
_x = (rec['start'] + rec['end']) / 2.0
_x, _y = self.get_prediction(series, band,
dates=np.array([_x]))
_x = _x[0][0]
_y = _y[0][0] + 250
axis.text(_x, _y, 'RMSE: %.3f' % rec['rmse'][band],
fontsize=18,
horizontalalignment='center')
elif desc == 'DOYPlot':
has_dates = all([r in self.yatsm_model.record.dtype.names
for r in ('spring_doy', 'autumn_doy')])
if self.config['calc_pheno'].value and has_dates:
colors = mpl.cm.Set1(np.linspace(0, 1, 9))[:, :-1]
color_cycle = itertools.cycle(colors)
for i, rec in enumerate(self.yatsm_model.record):
col = [c for c in color_cycle.next()]
artists.append(
axis.axvline(rec['spring_doy'], label='Model %i' % i,
c=col, lw=2)
)
axis.axvline(rec['autumn_doy'], label='Model %i' % i,
c=col, lw=2)
return artists
# RESULTS HELPER METHODS
def _fetch_results_saved(self):
""" Read YATSM results and return """
self.yatsm_model = MockResult()
row, col = self.series[0].py, self.series[0].px
data_cfg = {
'output': os.path.join(self.location,
self.config['results_folder'].value),
'output_prefix': (self.config['results_pattern'].value
.replace('*', ''))
}
result_filename = get_output_name(data_cfg, row)
logger.info('Attempting to open: {f}'.format(f=result_filename))
if not os.path.isfile(result_filename):
qgis_log('Could not find result for row {r} ({fn})'.format(
r=row, fn=result_filename))
return
z = np.load(result_filename)
if 'record' not in z.files:
raise KeyError('Cannot find "record" within saved result ({})'
.format(result_filename))
if 'metadata' not in z.files:
raise KeyError('Cannot find "metadata" within saved result ({})'
.format(result_filename))
metadata = z['metadata'].item()
if 'design' not in metadata['YATSM']:
raise KeyError('Cannot find "design" within saved result metadata '
'({})'.format(result_filename))
self._design = metadata['YATSM']['design_matrix']
self._design_info = metadata['YATSM']['design']
rec = z['record']
idx = np.where((rec['px'] == col) & (rec['py'] == row))[0]
self.yatsm_model.record = rec[idx]
def _fetch_results_live(self):
""" Run YATSM and get results """
logger.debug('Calculating YATSM results on the fly')
# Setup design matrix, Y, and dates
self.X = patsy.dmatrix(self.controls['design'].value,
{'x': self.series[0].images['ordinal'],
'sensor': self.series[0].sensor,
'pr': self.series[0].pathrow})
self._design_info = self.X.design_info.column_name_indexes
self.Y = self.series[0].data.astype(np.int16)
self.dates = np.asarray(self.series[0].images['ordinal'])
mask = self.Y[self.config['mask_band'].value[0] - 1, :]
Y_data = np.delete(self.Y, self.config['mask_band'].value[0] - 1,
axis=0)
# Mask out masked values
clear = np.in1d(mask, self.mask_values, invert=True)
valid = get_valid_mask(Y_data,
self.config['min_values'].value,
self.config['max_values'].value).astype(np.bool)
clear *= valid
# Setup parameters
estimator = sklearn.linear_model.Lasso(alpha=20)
reg = self.controls['regression_type'].value
if hasattr(yatsm.regression, 'packaged'):
if reg in yatsm.regression.packaged.packaged_regressions:
reg_fn = yatsm.regression.packaged.find_packaged_regressor(reg)
try:
estimator = jl.load(reg_fn)
except:
logger.error('Cannot load regressor: %s' % reg)
else:
logger.debug('Loaded regressor %s from %s' % (reg, reg_fn))
else:
logger.error('Cannot use unknown regression %s' % reg)
else:
logger.warning(
'Using failsafe Lasso(lambda=20) from scikit-learn. '
'Upgrade to yatsm>=0.5.1 to access more regressors.')
kwargs = dict(
estimator=estimator,
test_indices=self.controls['test_indices'].value,
consecutive=self.controls['consecutive'].value,
threshold=self.controls['threshold'].value,
min_obs=self.controls['min_obs'].value,
min_rmse=(None if self.controls['enable_min_rmse'].value else
self.controls['min_rmse'].value),
screening_crit=self.controls['screen_crit'].value,
remove_noise=self.controls['remove_noise'].value,
dynamic_rmse=self.controls['dynamic_rmse'].value,
)
self.yatsm_model = CCDCesque(**version_kwargs(kwargs))
# Don't want to have DEBUG logging when we run YATSM
log_level = logger.level
logger.setLevel(logging.INFO)
if self.controls['reverse'].value:
self.yatsm_model.fit(
np.flipud(self.X[clear, :]),
np.fliplr(Y_data[:, clear]),
self.dates[clear][::-1])
else:
self.yatsm_model.fit(
self.X[clear, :],
Y_data[:, clear],
self.dates[clear])
if self.controls['commit_test'].value:
self.yatsm_model.record = postprocess.commission_test(
self.yatsm_model, self.controls['commit_alpha'].value)
# if self.controls['robust_results'].value:
# self.coef_name = 'robust_coef'
# self.yatsm_model.record = postprocess.refit_record(
# self.yatsm_model, 'robust'
# else:
# self.coef_name = 'coef'
if self.config['calc_pheno'].value:
# TODO: parameterize band indices & scale factor
ltm = pheno.LongTermMeanPhenology()
self.yatsm_model.record = ltm.fit(self.yatsm_model)
# Restore log level
logger.setLevel(log_level)
# SETUP
def _init_metadata(self):
""" Setup metadata for series """
# Find MTL file
self.mtl_files = None
if self.config['metadata_file_pattern'].value:
search = find_files(
self.location, self.config['metadata_file_pattern'].value,
ignore_dirs=[self.config['results_folder'].value])
if len(search) == 0:
logger.error(
'Could not find image metadata with pattern {p}'.format(
p=self.config['metadata_file_pattern'].value))
if len(search) != len(self.series[0].images['date']):
logger.error(
'Inconsistent number of metadata files found: '
'{0} images vs {1} metadata files)'.format(
len(self.series[0].images['date']),
len(search)))
else:
self.mtl_files = search
# Setup metadata for series
self.series[0].metadata = ['sensor', 'pathrow', 'multitemp_screened']
self.series[0].metadata_names = ['Sensor', 'Path/Row',
'Multitemp Screened']
self.series[0].metadata_table = [False, False, False]
# Sensor ID
self.series[0].sensor = np.array([n[0:3] for n in
self.series[0].images['filename']])
# Path/row
self.series[0].pathrow = np.array([
'p{p}r{r}'.format(p=n[3:6], r=n[6:9]) for
n in self.series[0].images['filename']])
# Multitemporal noise screening - init to 0 (not screened)
# Will update this during model fitting
self.series[0].multitemp_screened = np.ones(self.series[0].n)
# Make an entry 0 so we get this in the unique values
self.series[0].multitemp_screened[0] = 0
# If we found MTL files, find cloud cover
if self.mtl_files is not None:
self.series[0].metadata.append('cloud_cover')
self.series[0].metadata_names.append('Cloud cover')
self.series[0].metadata_table.append(True)
self.series[0].cloud_cover = np.ones(self.series[0].n) * -9999
cloud_cover = {}
for mtl_file in self.mtl_files:
attrs = parse_landsat_MTL(mtl_file, ['LANDSAT_SCENE_ID',
'CLOUD_COVER'])
scene_ID = attrs.get('LANDSAT_SCENE_ID')
if scene_ID:
cloud_cover[scene_ID] = attrs.get('CLOUD_COVER', -9999.0)
for idx, _id in enumerate(self.series[0].images['id']):
self.series[0].cloud_cover[idx] = cloud_cover.get(_id, -9999.0)
if self.config['calc_pheno'].value:
self.series[0].metadata.append('pheno')
self.series[0].metadata_names.append('Phenology')
self.series[0].metadata_table.append(False)
# Initialize almost all as summer (SUM); first two as SPR/AUT
self.series[0].pheno = np.repeat('SUM', self.series[0].n)
self.series[0].pheno[0] = 'SPR'
self.series[0].pheno[1] = 'AUT'
class YATSMTimeSeries(timeseries_stacked.StackedTimeSeries):
""" Timeseries driver for CCDCesque algorithm implemented in YATSM
Requires a working installation of YATSM. For more information, visit
the [YATSM Github website](https://github.com/ceholden/yatsm).
This driver requires the following Python packages in addition to basic
TSTools package dependencies:
* [`scikit-learn`](http://scikit-learn.org/stable/)
* [`patsy`](https://patsy.readthedocs.org/en/latest/)
* [`yatsm`](https://github.com/ceholden/yatsm)
"""
description = 'YATSM CCDCesque Timeseries'
location = None
mask_values = np.array([2, 3, 4, 255])
has_results = True
# Driver configuration
config = OrderedDict((
('stack_pattern', ConfigItem('Stack pattern', 'L*stack')),
('date_index', ConfigItem('Date index', [9, 16])),
('date_format', ConfigItem('Date format', '%Y%j')),
('cache_folder', ConfigItem('Cache folder', 'cache')),
('results_folder', ConfigItem('Results folder', 'YATSM')),
('results_pattern', ConfigItem('Results pattern', 'yatsm_r*')),
('mask_band', ConfigItem('Mask band', [8])),
('min_values', ConfigItem('Min data values', [0])),
('max_values', ConfigItem('Max data values', [10000])),
('metadata_file_pattern',
ConfigItem('Metadata file pattern', 'L*MTL.txt')),
('calc_pheno', ConfigItem('LTM phenology', False)),
))
# Driver controls
controls_title = 'YATSM Algorithm Options'
controls = OrderedDict((
('calculate_live', ConfigItem('Calculate live', True)),
('consecutive', ConfigItem('Consecutive', 5)),
('min_obs', ConfigItem('Min obs.', 16)),
('threshold', ConfigItem('Threshold', 4.0)),
('enable_min_rmse', ConfigItem('Use min RMSE?', True)),
('min_rmse', ConfigItem('Min RMSE', 100.0)),
('design', ConfigItem('Design', '1 + x + harm(x, 1)')),
('test_indices', ConfigItem('Test indices', np.array([2, 3, 4, 5]))),
('dynamic_rmse', ConfigItem('Dynamic RMSE', True)),
('screen_crit', ConfigItem('Screening crit value', 400.0)),
('remove_noise', ConfigItem('Remove noise', True)),
('reverse', ConfigItem('Reverse', False)),
('regression_type', ConfigItem('Regression type', 'sklearn_Lasso20')),
('robust_results', ConfigItem('Robust results', False)),
('commit_test', ConfigItem('Commission test', False)),
('commit_alpha', ConfigItem('Commission test alpha', 0.10)),
))
def __init__(self, location, config=None):
super(YATSMTimeSeries, self).__init__(location, config=config)
# Check for YATSM imports
if not has_yatsm:
raise ImportError(has_yatsm_msg)
if self.config['calc_pheno'].value and not has_yatsm_pheno:
raise ImportError(has_yatsm_pheno_msg)
# Find extra metadata
self._init_metadata()
# Setup YATSM
self.yatsm_model = None
self.X = None
self.Y = None
self.coef_name = 'coef'
# Setup min/max values
desc, _min_values = self.config['min_values']
if len(_min_values) == 1:
_min_values = np.repeat(_min_values, self.series[0].count - 1)
self.config['min_values'] = ConfigItem(desc, _min_values)
desc, _max_values = self.config['max_values']
if len(_max_values) == 1:
_max_values = np.repeat(_max_values, self.series[0].count - 1)
self.config['max_values'] = ConfigItem(desc, _max_values)
def set_custom_controls(self, values):
logger.debug('Setting custom values')
for val, attr in zip(values, self.controls):
desc, current_val = self.controls[attr]
if isinstance(val, type(current_val)):
self.controls[attr] = ConfigItem(desc, val)
else:
# Make an exception for minimum RMSE since we can pass None
if attr == 'min_rmse' and isinstance(val, float):
self.controls[attr] = ConfigItem(desc, val)
else:
msg = 'Could not set {k} to {v} (current: {c})'.format(
k=attr, v=val, c=current_val)
raise ValueError(msg)
def fetch_results(self):
""" Read or calculate results for current pixel """
if self.controls['calculate_live'].value:
self._fetch_results_live()
else:
self._fetch_results_saved()
# Update multitemporal screening metadata
if self.yatsm_model:
if (self.controls['calculate_live']
and hasattr(self.yatsm_model, 'X')):
self.series[0].multitemp_screened = \
np.in1d(self.X[:, 1], self.yatsm_model.X[:, 1],
invert=True).astype(np.uint8)
if self.config['calc_pheno'].value:
for rec in self.yatsm_model.record:
# Find dates in record
idx = np.where(
(self.series[0].images['ordinal'] >= rec['start'])
& (self.series[0].images['ordinal'] <= rec['end']))[0]
# Put observations into SPR/SUM/AUT
_spr = np.where(
self.series[0].images['doy'][idx] <= rec['spring_doy']
)[0]
_sum = np.where(
(self.series[0].images['doy'][idx] > rec['spring_doy'])
& (self.series[0].images['doy'][idx] <
rec['autumn_doy']))[0]
_aut = np.where(
self.series[0].images['doy'][idx] >= rec['autumn_doy']
)[0]
self.series[0].pheno[idx[_spr]] = 'SPR'
self.series[0].pheno[idx[_sum]] = 'SUM'
self.series[0].pheno[idx[_aut]] = 'AUT'
def get_prediction(self, series, band, dates=None):
""" Return prediction for a given band
Args:
series (int): index of Series used for prediction
band (int): index of band to return
dates (iterable): list or np.ndarray of ordinal dates to predict; if
None, predicts for every date within timeseries (default: None)
Returns:
iterable: sequence of tuples (1D NumPy arrays, x and y) containing
predictions
"""
if series > 0:
return
if self.yatsm_model is None or len(self.yatsm_model.record) == 0:
return
if band >= self.yatsm_model.record[self.coef_name].shape[2]:
logger.debug('Not results for band %i' % band)
return
# Setup output
mx = []
my = []
# Don't predict with any categorical information
eqn = (self.controls['design'].value
if self.controls['calculate_live'].value else self._design)
design = re.sub(r'[\+\-][\ ]+C\(.*\)', '', eqn)
coef_columns = []
for k, v in self._design_info.iteritems():
if not re.match('C\(.*\)', k):
coef_columns.append(v)
coef_columns = np.sort(np.asarray(coef_columns))
for rec in self.yatsm_model.record:
# Check for reverse
if rec['end'] < rec['start']:
i_step = -1
else:
i_step = 1
# Date range to predict
if dates is not None:
end = max(rec['break'], rec['end'])
_mx = dates[np.where((dates >= rec['start'])
& (dates <= end))[0]]
else:
_mx = np.arange(rec['start'], rec['end'], i_step)
if _mx.size == 0:
continue
# Coefficients to use for prediction
_coef = rec[self.coef_name][coef_columns, band]
# Setup design matrix
_mX = patsy.dmatrix(design, {'x': _mx}).T
# Predict
_my = np.dot(_coef, _mX)
# Transform ordinal back to datetime for plotting
_mx = np.array([dt.fromordinal(int(_x)) for _x in _mx])
mx.append(_mx)
my.append(_my)
return mx, my
def get_breaks(self, series, band):
""" Return break points for a given band
Args:
series (int): index of Series for prediction
band (int): index of band to return
Returns:
iterable: sequence of tuples (1D NumPy arrays, x and y) containing
break points
"""
if self.yatsm_model is None:
return
# Setup output
bx = []
by = []
if len(self.yatsm_model.record) > 0:
for rec in self.yatsm_model.record:
if rec['break'] != 0:
_bx = dt.fromordinal(int(rec['break']))
index = np.where(
self.series[series].images['date'] == _bx)[0]
if (index.size > 0
and index[0] < self.series[series].data.shape[1]):
bx.append(_bx)
by.append(self.series[series].data[band, index[0]])
else:
logger.warning('Could not determine breakpoint')
return bx, by
def get_residuals(self, series, band):
""" Return model residuals (y - predicted yhat) for a given band
Args:
series (int): index of Series for residuals
band (int): index of band to return
Returns:
iterable: sequence of tuples (1D NumPy arrays, x and y) containing
residual dates and values
"""
if self.yatsm_model is None:
return
rx, ry = [], []
X, y = self.get_data(series, band, mask=settings.plot['mask'])
predict = self.get_prediction(series, band, dates=X['ordinal'])
if predict is None:
return
date, yhat = predict
for _date, _yhat in zip(date, yhat):
idx = np.in1d(X['date'], _date)
resid = y[idx] - _yhat
rx.append(_date)
ry.append(resid)
return rx, ry
def get_plot(self, series, band, axis, desc):
""" Plot some information on an axis for a plot of some description
Args:
series (int): index of Series for residuals
band (int): index of band to return
axis (matplotlib.axes._subplots.Axes): a matplotlib axis to plot on
desc (str): description of plot, usually a plot class from
`tstools.plots`
Returns:
iterable: list of artists to include in legend
"""
artists = []
if desc == 'TSPlot':
for rec in self.yatsm_model.record:
_x = (rec['start'] + rec['end']) / 2.0
_x, _y = self.get_prediction(series,
band,
dates=np.array([_x]))
_x = _x[0][0]
_y = _y[0][0] + 250
axis.text(_x,
_y,
'RMSE: %.3f' % rec['rmse'][band],
fontsize=18,
horizontalalignment='center')
elif desc == 'DOYPlot':
has_dates = all([
r in self.yatsm_model.record.dtype.names
for r in ('spring_doy', 'autumn_doy')
])
if self.config['calc_pheno'].value and has_dates:
colors = mpl.cm.Set1(np.linspace(0, 1, 9))[:, :-1]
color_cycle = itertools.cycle(colors)
for i, rec in enumerate(self.yatsm_model.record):
col = [c for c in color_cycle.next()]
artists.append(
axis.axvline(rec['spring_doy'],
label='Model %i' % i,
c=col,
lw=2))
axis.axvline(rec['autumn_doy'],
label='Model %i' % i,
c=col,
lw=2)
return artists
# RESULTS HELPER METHODS
def _fetch_results_saved(self):
""" Read YATSM results and return """
self.yatsm_model = MockResult()
row, col = self.series[0].py, self.series[0].px
data_cfg = {
'output':
os.path.join(self.location, self.config['results_folder'].value),
'output_prefix':
(self.config['results_pattern'].value.replace('*', ''))
}
result_filename = get_output_name(data_cfg, row)
logger.info('Attempting to open: {f}'.format(f=result_filename))
if not os.path.isfile(result_filename):
qgis_log('Could not find result for row {r} ({fn})'.format(
r=row, fn=result_filename))
return
z = np.load(result_filename)
if 'record' not in z.files:
raise KeyError(
'Cannot find "record" within saved result ({})'.format(
result_filename))
if 'metadata' not in z.files:
raise KeyError(
'Cannot find "metadata" within saved result ({})'.format(
result_filename))
metadata = z['metadata'].item()
if 'design' not in metadata['YATSM']:
raise KeyError('Cannot find "design" within saved result metadata '
'({})'.format(result_filename))
self._design = metadata['YATSM']['design_matrix']
self._design_info = metadata['YATSM']['design']
rec = z['record']
idx = np.where((rec['px'] == col) & (rec['py'] == row))[0]
self.yatsm_model.record = rec[idx]
def _fetch_results_live(self):
""" Run YATSM and get results """
logger.debug('Calculating YATSM results on the fly')
# Setup design matrix, Y, and dates
self.X = patsy.dmatrix(
self.controls['design'].value, {
'x': self.series[0].images['ordinal'],
'sensor': self.series[0].sensor,
'pr': self.series[0].pathrow
})
self._design_info = self.X.design_info.column_name_indexes
self.Y = self.series[0].data.astype(np.int16)
self.dates = np.asarray(self.series[0].images['ordinal'])
mask = self.Y[self.config['mask_band'].value[0] - 1, :]
Y_data = np.delete(self.Y,
self.config['mask_band'].value[0] - 1,
axis=0)
# Mask out masked values
clear = np.in1d(mask, self.mask_values, invert=True)
valid = get_valid_mask(Y_data, self.config['min_values'].value,
self.config['max_values'].value).astype(np.bool)
clear *= valid
# Setup parameters
estimator = sklearn.linear_model.Lasso(alpha=20)
reg = self.controls['regression_type'].value
if hasattr(yatsm.regression, 'packaged'):
if reg in yatsm.regression.packaged.packaged_regressions:
reg_fn = yatsm.regression.packaged.find_packaged_regressor(reg)
try:
estimator = jl.load(reg_fn)
except:
logger.error('Cannot load regressor: %s' % reg)
else:
logger.debug('Loaded regressor %s from %s' % (reg, reg_fn))
else:
logger.error('Cannot use unknown regression %s' % reg)
else:
logger.warning(
'Using failsafe Lasso(lambda=20) from scikit-learn. '
'Upgrade to yatsm>=0.5.1 to access more regressors.')
kwargs = dict(
estimator=estimator,
test_indices=self.controls['test_indices'].value,
consecutive=self.controls['consecutive'].value,
threshold=self.controls['threshold'].value,
min_obs=self.controls['min_obs'].value,
min_rmse=(None if self.controls['enable_min_rmse'].value else
self.controls['min_rmse'].value),
screening_crit=self.controls['screen_crit'].value,
remove_noise=self.controls['remove_noise'].value,
dynamic_rmse=self.controls['dynamic_rmse'].value,
)
self.yatsm_model = CCDCesque(**version_kwargs(kwargs))
# Don't want to have DEBUG logging when we run YATSM
log_level = logger.level
logger.setLevel(logging.INFO)
if self.controls['reverse'].value:
self.yatsm_model.fit(np.flipud(self.X[clear, :]),
np.fliplr(Y_data[:, clear]),
self.dates[clear][::-1])
else:
self.yatsm_model.fit(self.X[clear, :], Y_data[:, clear],
self.dates[clear])
if self.controls['commit_test'].value:
self.yatsm_model.record = postprocess.commission_test(
self.yatsm_model, self.controls['commit_alpha'].value)
# if self.controls['robust_results'].value:
# self.coef_name = 'robust_coef'
# self.yatsm_model.record = postprocess.refit_record(
# self.yatsm_model, 'robust'
# else:
# self.coef_name = 'coef'
if self.config['calc_pheno'].value:
# TODO: parameterize band indices & scale factor
ltm = pheno.LongTermMeanPhenology()
self.yatsm_model.record = ltm.fit(self.yatsm_model)
# Restore log level
logger.setLevel(log_level)
# SETUP
def _init_metadata(self):
""" Setup metadata for series """
# Find MTL file
self.mtl_files = None
if self.config['metadata_file_pattern'].value:
search = find_files(
self.location,
self.config['metadata_file_pattern'].value,
ignore_dirs=[self.config['results_folder'].value])
if len(search) == 0:
logger.error(
'Could not find image metadata with pattern {p}'.format(
p=self.config['metadata_file_pattern'].value))
if len(search) != len(self.series[0].images['date']):
logger.error('Inconsistent number of metadata files found: '
'{0} images vs {1} metadata files)'.format(
len(self.series[0].images['date']),
len(search)))
else:
self.mtl_files = search
# Setup metadata for series
self.series[0].metadata = ['sensor', 'pathrow', 'multitemp_screened']
self.series[0].metadata_names = [
'Sensor', 'Path/Row', 'Multitemp Screened'
]
self.series[0].metadata_table = [False, False, False]
# Sensor ID
self.series[0].sensor = np.array(
[n[0:3] for n in self.series[0].images['filename']])
# Path/row
self.series[0].pathrow = np.array([
'p{p}r{r}'.format(p=n[3:6], r=n[6:9])
for n in self.series[0].images['filename']
])
# Multitemporal noise screening - init to 0 (not screened)
# Will update this during model fitting
self.series[0].multitemp_screened = np.ones(self.series[0].n)
# Make an entry 0 so we get this in the unique values
self.series[0].multitemp_screened[0] = 0
# If we found MTL files, find cloud cover
if self.mtl_files is not None:
self.series[0].metadata.append('cloud_cover')
self.series[0].metadata_names.append('Cloud cover')
self.series[0].metadata_table.append(True)
self.series[0].cloud_cover = np.ones(self.series[0].n) * -9999
cloud_cover = {}
for mtl_file in self.mtl_files:
attrs = parse_landsat_MTL(mtl_file,
['LANDSAT_SCENE_ID', 'CLOUD_COVER'])
scene_ID = attrs.get('LANDSAT_SCENE_ID')
if scene_ID:
cloud_cover[scene_ID] = attrs.get('CLOUD_COVER', -9999.0)
for idx, _id in enumerate(self.series[0].images['id']):
self.series[0].cloud_cover[idx] = cloud_cover.get(_id, -9999.0)
if self.config['calc_pheno'].value:
self.series[0].metadata.append('pheno')
self.series[0].metadata_names.append('Phenology')
self.series[0].metadata_table.append(False)
# Initialize almost all as summer (SUM); first two as SPR/AUT
self.series[0].pheno = np.repeat('SUM', self.series[0].n)
self.series[0].pheno[0] = 'SPR'
self.series[0].pheno[1] = 'AUT'
class YATSMTimeSeries(timeseries_stacked.StackedTimeSeries):
""" Timeseries driver for YATSM algorithm
"""
description = 'YATSM CCDCesque Timeseries'
location = None
mask_values = np.array([2, 3, 4, 255])
# Driver configuration
_stack_pattern = 'L*stack'
_date_index = [9, 16]
_date_format = '%Y%j'
_cache_folder = 'cache'
_results_folder = 'YATSM'
_results_pattern = 'yatsm_r*'
_mask_band = [8]
_min_values = [0]
_max_values = [10000]
_metadata_file_pattern = 'L*MTL.txt'
_calc_pheno = False
config = ['_stack_pattern',
'_date_index',
'_date_format',
'_cache_folder',
'_results_folder',
'_results_pattern',
'_mask_band',
'_min_values', '_max_values',
'_metadata_file_pattern',
'_calc_pheno']
config_names = [
'Stack pattern',
'Date index',
'Date format',
'Cache folder',
'Results folder',
'Results pattern',
'Mask band',
'Min data values', 'Max data values',
'Metadata file pattern',
'LTM phenology']
# Driver controls
_calculate_live = True
_consecutive = 5
_min_obs = 16
_threshold = 4.0
_enable_min_rmse = True
_min_rmse = 100
_design = '1 + x + harm(x, 1)'
_test_indices = np.array([2, 3, 4, 5])
_dynamic_rmse = True
_screen_crit = 400.0
_remove_noise = True
_reverse = False
_robust_results = False
_commit_test = False
_commit_alpha = 0.01
# Requires YATSM>=v0.5.0
_regression_type = 'sklearn_Lasso20'
controls_title = 'YATSM Algorithm Options'
controls = [
'_calculate_live',
'_consecutive',
'_min_obs',
'_threshold',
'_enable_min_rmse',
'_min_rmse',
'_design',
'_test_indices',
'_dynamic_rmse',
'_screen_crit',
'_remove_noise',
'_reverse',
'_regression_type',
'_robust_results',
'_commit_test',
'_commit_alpha']
controls_names = [
'Calculate live',
'Consecutive',
'Min Observations',
'Threshold',
'Use min RMSE?',
'Min RMSE',
'Design',
'Test indices',
'Dynamic RMSE',
'Screening critical value',
'Remove noise',
'Run in reverse',
'Regression type',
'Robust results',
'Commission test',
'Commission test alpha']
def __init__(self, location, config=None):
super(YATSMTimeSeries, self).__init__(location, config=config)
# Check for YATSM imports
self._check_yatsm()
# Find extra metadata
self._init_metadata()
# Setup YATSM
self.yatsm_model = None
self.X = None
self.Y = None
self.coef_name = 'coef'
# Setup min/max values
if len(self._min_values) == 1:
self._min_values = self._min_values * (self.series[0].count - 1)
if len(self._max_values) == 1:
self._max_values = self._max_values * (self.series[0].count - 1)
self._min_values = np.asarray(self._min_values)
self._max_values = np.asarray(self._max_values)
def set_custom_controls(self, values):
logger.debug('Setting custom values')
for v, k in zip(values, self.controls):
current_value = getattr(self, k)
if isinstance(v, type(current_value)):
setattr(self, k, v)
else:
# Make an exception for minimum RMSE since we can pass None
if k == 'min_rmse' and isinstance(v, float):
setattr(self, k, v)
else:
msg = 'Could not set {k} to {v} (current: {c})'.format(
k=k, v=v, c=current_value)
raise Exception(msg)
def fetch_results(self):
""" Read or calculate results for current pixel """
if self._calculate_live:
self._fetch_results_live()
else:
self._fetch_results_saved()
# Update multitemporal screening metadata
if self.yatsm_model:
self.series[0].multitemp_screened = \
np.in1d(self.X[:, 1], self.yatsm_model.X[:, 1],
invert=True).astype(np.uint8)
if self._calc_pheno:
for rec in self.yatsm_model.record:
# Find dates in record
idx = np.where(
(self.series[0].images['ordinal'] >= rec['start']) &
(self.series[0].images['ordinal'] <= rec['end']))[0]
# Put observations into SPR/SUM/AUT
_spr = np.where(self.series[0].images['doy'][idx] <=
rec['spring_doy'])[0]
_sum = np.where((self.series[0].images['doy'][idx] >
rec['spring_doy']) &
(self.series[0].images['doy'][idx] <
rec['autumn_doy']))[0]
_aut = np.where(self.series[0].images['doy'][idx] >=
rec['autumn_doy'])[0]
self.series[0].pheno[idx[_spr]] = 'SPR'
self.series[0].pheno[idx[_sum]] = 'SUM'
self.series[0].pheno[idx[_aut]] = 'AUT'
def get_prediction(self, series, band, dates=None):
""" Return prediction for a given band
Args:
series (int): index of Series used for prediction
band (int): index of band to return
dates (iterable): list or np.ndarray of ordinal dates to predict; if
None, predicts for every date within timeseries (default: None)
Returns:
iterable: sequence of tuples (1D NumPy arrays, x and y) containing
predictions
"""
if series > 0:
return
if self.yatsm_model is None or len(self.yatsm_model.record) == 0:
return
if band >= self.yatsm_model.record[self.coef_name].shape[2]:
logger.debug('Not results for band %i' % band)
return
# Setup output
mx = []
my = []
# Don't predict with any categorical information
design = re.sub(r'[\+\-][\ ]+C\(.*\)', '', self._design)
coef_columns = []
for k, v in self._design_info.column_name_indexes.iteritems():
if not re.match('C\(.*\)', k):
coef_columns.append(v)
coef_columns = np.asarray(coef_columns)
for rec in self.yatsm_model.record:
# Check for reverse
if rec['end'] < rec['start']:
i_step = -1
else:
i_step = 1
# Date range to predict
if dates is not None:
end = max(rec['break'], rec['end'])
_mx = dates[np.where((dates >= rec['start']) &
(dates <= end))[0]]
else:
_mx = np.arange(rec['start'], rec['end'], i_step)
if _mx.size == 0:
continue
# Coefficients to use for prediction
_coef = rec[self.coef_name][coef_columns, band]
# Setup design matrix
_mX = patsy.dmatrix(design, {'x': _mx}).T
# Predict
_my = np.dot(_coef, _mX)
# Transform ordinal back to datetime for plotting
_mx = np.array([dt.fromordinal(int(_x)) for _x in _mx])
mx.append(_mx)
my.append(_my)
return mx, my
def get_breaks(self, series, band):
""" Return break points for a given band
Args:
series (int): index of Series for prediction
band (int): index of band to return
Returns:
iterable: sequence of tuples (1D NumPy arrays, x and y) containing
break points
"""
if self.yatsm_model is None:
return
# Setup output
bx = []
by = []
if len(self.yatsm_model.record) > 0:
for rec in self.yatsm_model.record:
if rec['break'] != 0:
_bx = dt.fromordinal(int(rec['break']))
index = np.where(self.series[series].images['date']
== _bx)[0]
if (index.size > 0 and
index[0] < self.series[series].data.shape[1]):
bx.append(_bx)
by.append(self.series[series].data[band, index[0]])
else:
logger.warning('Could not determine breakpoint')
return bx, by
def get_residuals(self, series, band):
""" Return model residuals (y - predicted yhat) for a given band
Args:
series (int): index of Series for residuals
band (int): index of band to return
Returns:
iterable: sequence of tuples (1D NumPy arrays, x and y) containing
residual dates and values
"""
if self.yatsm_model is None:
return
rx, ry = [], []
X, y = self.get_data(series, band, mask=settings.plot['mask'])
predict = self.get_prediction(series, band, dates=X['ordinal'])
if predict is None:
return
date, yhat = predict
for _date, _yhat in zip(date, yhat):
idx = np.in1d(X['date'], _date)
resid = y[idx] - _yhat
rx.append(_date)
ry.append(resid)
return rx, ry
def get_plot(self, series, band, axis, desc):
""" Plot some information on an axis for a plot of some description
Args:
series (int): index of Series for residuals
band (int): index of band to return
axis (matplotlib.axes._subplots.Axes): a matplotlib axis to plot on
desc (str): description of plot, usually a plot class from
`tstools.plots`
Returns:
iterable: list of artists to include in legend
"""
artists = []
if desc == 'TSPlot':
for rec in self.yatsm_model.record:
_x = (rec['start'] + rec['end']) / 2.0
_x, _y = self.get_prediction(series, band,
dates=np.array([_x]))
_x = _x[0][0]
_y = _y[0][0] + 250
axis.text(_x, _y, 'RMSE: %.3f' % rec['rmse'][band],
fontsize=18,
horizontalalignment='center')
elif desc == 'DOYPlot':
names = self.yatsm_model.record.dtype.names
if self._calc_pheno and all([r in names for r in
('spring_doy', 'autumn_doy')]):
colors = mpl.cm.Set1(np.linspace(0, 1, 9))[:, :-1]
color_cycle = itertools.cycle(colors)
for i, rec in enumerate(self.yatsm_model.record):
col = [c for c in color_cycle.next()]
artists.append(
axis.axvline(rec['spring_doy'], label='Model %i' % i,
c=col, lw=2)
)
axis.axvline(rec['autumn_doy'], label='Model %i' % i,
c=col, lw=2)
return artists
# RESULTS HELPER METHODS
def _fetch_results_saved(self):
""" Read YATSM results and return """
raise NotImplementedError('No saved results reading just yet...')
def _fetch_results_live(self):
""" Run YATSM and get results """
logger.debug('Calculating YATSM results on the fly')
# Setup design matrix, Y, and dates
self.X = patsy.dmatrix(self._design,
{
'x': self.series[0].images['ordinal'],
'sensor': self.series[0].sensor,
'pr': self.series[0].pathrow
})
self._design_info = self.X.design_info
self.Y = self.series[0].data.astype(np.int16)
self.dates = np.asarray(self.series[0].images['ordinal'])
mask = self.Y[self._mask_band[0] - 1, :]
Y_data = np.delete(self.Y, self._mask_band[0] - 1, axis=0)
# Mask out masked values
clear = np.in1d(mask, self.mask_values, invert=True)
valid = get_valid_mask(Y_data,
self._min_values,
self._max_values).astype(np.bool)
clear *= valid
# Setup Y
# Setup parameters
lm = sklearn.linear_model.Lasso(alpha=20)
reg = self._regression_type
print(self._regression_type)
if hasattr(yatsm.regression, 'packaged'):
if reg in yatsm.regression.packaged.packaged_regressions:
reg_fn = yatsm.regression.packaged.find_packaged_regressor(reg)
try:
lm = jl.load(reg_fn)
except:
logger.error('Cannot load regressor: %s' % reg)
else:
logger.debug('Loaded regressor %s from %s' % (reg, reg_fn))
else:
logger.error('Cannot use unknown regression %s' % reg)
else:
logger.warning('Using failsafe Lasso(lambda=20) from scikit-learn. '
'Upgrade to yatsm>=0.5.1 to access more regressors.')
kwargs = dict(
test_indices=self._test_indices,
consecutive=self._consecutive,
threshold=self._threshold,
min_obs=self._min_obs,
min_rmse=None if self._enable_min_rmse else self._min_rmse,
screening_crit=self._screen_crit,
remove_noise=self._remove_noise,
dynamic_rmse=self._dynamic_rmse,
)
self.yatsm_model = CCDCesque(lm=lm, **kwargs)
# Don't want to have DEBUG logging when we run YATSM
log_level = logger.level
logger.setLevel(logging.INFO)
if self._reverse:
self.yatsm_model.fit(
np.flipud(self.X[clear, :]),
np.fliplr(Y_data[:, clear]),
np.fliplr(self.dates[clear]))
else:
self.yatsm_model.fit(
self.X[clear, :],
Y_data[:, clear],
self.dates[clear])
if self._commit_test:
self.yatsm_model.record = postprocess.commission_test(
self.yatsm_model, self._commit_alpha)
# if self._robust_results:
# self.coef_name = 'robust_coef'
# self.yatsm_model.record = postprocess.refit_record(
# self.yatsm_model, 'robust'
# else:
# self.coef_name = 'coef'
if self._calc_pheno:
# TODO: parameterize band indices & scale factor
ltm = pheno.LongTermMeanPhenology(self.yatsm_model)
self.yatsm_model.record = ltm.fit()
# Restore log level
logger.setLevel(log_level)
# SETUP
def _init_metadata(self):
""" Setup metadata for series """
# Find MTL file
self.mtl_files = None
if self._metadata_file_pattern:
search = find_files(self.location, self._metadata_file_pattern,
ignore_dirs=[self._results_folder])
if len(search) == 0:
logger.error(
'Could not find image metadata with pattern {p}'.format(
p=self._metadata_file_pattern))
if len(search) != len(self.series[0].images['date']):
logger.error('Inconsistent number of metadata files found: '
'{0} images vs {1} metadata files)'.format(
len(self.series[0].images['date']),
len(search)))
else:
self.mtl_files = search
# Setup metadata for series
self.series[0].metadata = ['sensor', 'pathrow', 'multitemp_screened']
self.series[0].metadata_names = ['Sensor', 'Path/Row',
'Multitemp Screened']
self.series[0].metadata_table = [False, False, False]
# Sensor ID
self.series[0].sensor = np.array([n[0:3] for n in
self.series[0].images['filename']])
# Path/row
self.series[0].pathrow = np.array([
'p{p}r{r}'.format(p=n[3:6], r=n[6:9]) for
n in self.series[0].images['filename']])
# Multitemporal noise screening - init to 0 (not screened)
# Will update this during model fitting
self.series[0].multitemp_screened = np.ones(self.series[0].n)
# Make an entry 0 so we get this in the unique values
self.series[0].multitemp_screened[0] = 0
# If we found MTL files, find cloud cover
if self.mtl_files is not None:
self.series[0].metadata.append('cloud_cover')
self.series[0].metadata_names.append('Cloud cover')
self.series[0].metadata_table.append(True)
self.series[0].cloud_cover = np.zeros(self.series[0].n)
for i, mtl_file in enumerate(self.mtl_files):
self.series[0].cloud_cover[i] = parse_landsat_MTL(
mtl_file, 'CLOUD_COVER')
if self._calc_pheno:
self.series[0].metadata.append('pheno')
self.series[0].metadata_names.append('Phenology')
self.series[0].metadata_table.append(False)
# Initialize almost all as summer (SUM); first two as SPR/AUT
self.series[0].pheno = np.repeat('SUM', self.series[0].n)
self.series[0].pheno[0] = 'SPR'
self.series[0].pheno[1] = 'AUT'
def _check_yatsm(self):
""" Check if YATSM is available
"""
try:
global yatsm
global CCDCesque, postprocess
global harm
global get_valid_mask
import yatsm
from yatsm.algorithms import CCDCesque, postprocess
from yatsm._cyprep import get_valid_mask
from yatsm.regression.transforms import harm
except ImportError as e:
raise ImportError('Could not import YATSM because it could not '
'import a dependency (%s)' % e.message)
except:
raise ImportError('Could not import YATSM for an unknown reason '
'(%s)' % e.message)
else:
self.has_results = True
if self._calc_pheno:
try:
global pheno
import yatsm.phenology as pheno
except:
msg = ('Could not import YATSM phenology module. '
'Make sure you have R and rpy2 installed.')
raise ImportError(msg)
def _fetch_results_live(self):
""" Run YATSM and get results """
logger.debug('Calculating YATSM results on the fly')
# Setup design matrix, Y, and dates
self.X = patsy.dmatrix(self._design,
{
'x': self.series[0].images['ordinal'],
'sensor': self.series[0].sensor,
'pr': self.series[0].pathrow
})
self._design_info = self.X.design_info
self.Y = self.series[0].data.astype(np.int16)
self.dates = np.asarray(self.series[0].images['ordinal'])
mask = self.Y[self._mask_band[0] - 1, :]
Y_data = np.delete(self.Y, self._mask_band[0] - 1, axis=0)
# Mask out masked values
clear = np.in1d(mask, self.mask_values, invert=True)
valid = get_valid_mask(Y_data,
self._min_values,
self._max_values).astype(np.bool)
clear *= valid
# Setup Y
# Setup parameters
lm = sklearn.linear_model.Lasso(alpha=20)
reg = self._regression_type
print(self._regression_type)
if hasattr(yatsm.regression, 'packaged'):
if reg in yatsm.regression.packaged.packaged_regressions:
reg_fn = yatsm.regression.packaged.find_packaged_regressor(reg)
try:
lm = jl.load(reg_fn)
except:
logger.error('Cannot load regressor: %s' % reg)
else:
logger.debug('Loaded regressor %s from %s' % (reg, reg_fn))
else:
logger.error('Cannot use unknown regression %s' % reg)
else:
logger.warning('Using failsafe Lasso(lambda=20) from scikit-learn. '
'Upgrade to yatsm>=0.5.1 to access more regressors.')
kwargs = dict(
test_indices=self._test_indices,
consecutive=self._consecutive,
threshold=self._threshold,
min_obs=self._min_obs,
min_rmse=None if self._enable_min_rmse else self._min_rmse,
screening_crit=self._screen_crit,
remove_noise=self._remove_noise,
dynamic_rmse=self._dynamic_rmse,
)
self.yatsm_model = CCDCesque(lm=lm, **kwargs)
# Don't want to have DEBUG logging when we run YATSM
log_level = logger.level
logger.setLevel(logging.INFO)
if self._reverse:
self.yatsm_model.fit(
np.flipud(self.X[clear, :]),
np.fliplr(Y_data[:, clear]),
np.fliplr(self.dates[clear]))
else:
self.yatsm_model.fit(
self.X[clear, :],
Y_data[:, clear],
self.dates[clear])
if self._commit_test:
self.yatsm_model.record = postprocess.commission_test(
self.yatsm_model, self._commit_alpha)
# if self._robust_results:
# self.coef_name = 'robust_coef'
# self.yatsm_model.record = postprocess.refit_record(
# self.yatsm_model, 'robust'
# else:
# self.coef_name = 'coef'
if self._calc_pheno:
# TODO: parameterize band indices & scale factor
ltm = pheno.LongTermMeanPhenology(self.yatsm_model)
self.yatsm_model.record = ltm.fit()
# Restore log level
logger.setLevel(log_level)
